An approval gate (also called a human checkpoint) is a deliberate pause point in an AI coding agent's execution where the agent stops, surfaces its current state, and waits for human confirmation before continuing. Most developers run zero gates and absorb the cost when something goes sideways. The opposite failure — approval prompts on every tool call — just rebuilds a slow human workflow. This tutorial shows you where the three minimum effective gates are, what belongs at each one, and how to implement them with patterns you can copy directly into your project. TL;DR Three gates cover the majority of meaningful risk without meaningful overhead: Plan review gate — approve the agent's approach before it touches any files Findings review gate — confirm what the agent discovered before it acts on it Diff-before-push gate — inspect the full diff before any code leaves your machine All three are implementable today using CLAUDE.md prompts and a shell function. No specialized tooling required.…